Generation and Dramatization of Detective Stories

Transcription

1 SBC Journal on Interactve Systems, volume 5, number 2, Generaton and Dramatzaton of Detectve Stores Smone D.J. Barbosa, Edrle Soares de Lma, Antono L. Furtado, Bruno Fejó PUC-Ro Departamento de Informátca Ro de Janero, RJ Brazl {smone, elma, furtado, Abstract We ntroduce n the present paper an operatonally defned subclass wthn the genre of detectve stores, specfed on the bass of the logc programmng model adopted n our Logtell project. Specal attenton s gven to the treatment of communcatve events. An SWI-Prolog plan-based tool was developed to compose consstent plots, conformng to the conventons of the genre. Seven crmnal cases generated by the tool are descrbed as llustraton. It s shown, usng the PlotBoard nterface, how to run plan-based composton n nteractve stepwse mode. We have also developed a storytellng system capable of representng the stores n the format of nteractve comc books on tablet computers. Keywords Dgtal Entertanment; Plot Composton; Communcatve Speech Acts; Detectve Stores; Logc Programmng; Plan Generaton; Plan Recognton; Comc Books. I. INTRODUCTION The basc objectve of ths paper s to propose a strategy to defne a subclass of the genre of detectve stores, on the bass of a logc-programmng model ntroduced n [1] and more rgorously formulated n [2]. The defnton specfes: 1. what can exst n a state of the underlyng mn-world 2. how states can be changed, and 3. the factors drvng the characters to act. Informally speakng, such threefold specfcatons determne what knd of facts can hold n each state, what events (denoted by operatons defned n terms of precondtons and post-condtons) can change a state, and what goals can motvate the actng characters to trgger events that compose consstent detectve story plots. In turn, the composton process can be effectuated by a plan-generator algorthm. What plots wll be generated depends not only on how the genre was specfed, but also on the specfcaton of a sutable ntal state, n whch the entty nstances (ncludng characters and objects) that wll fgure n the plots are ntroduced, together wth ther ntal propertes (attrbute-values and relatonshp connectons wth other entty nstances). To accommodate characters wth dstnct personalty trats, and to deal wth a dversty of pecular or even anomalous stuatons, we started to allow the generc pre-condtons and post-condtons of the event-producng operatons to be complemented by what we call condtoner clauses, as one addtonal part of the ntal state specfcatons. Choosng dfferent ntal states s thus a way to acheve a vared repertore wthn a gven genre whch s further enlarged by both () the ablty of the plan-generator to backtrack to produce dfferent plans to reach the same goals and () the user-nteracton features that can be bult nto the algorthm. One may object to the use of plan-generaton to help composng stores, even wth the user's nteracton, on the grounds that the resultng plots would be essentally predctable. To ths we would reply that, whle never pretendng to equal the creatvty of talented wrters, planners gven ther ablty to systematcally explore the consequences of rule-based specfcatons can often come up wth the unexpected, as wll be exemplfed at the end of part A of Secton II. Furthermore, some degree of uncertanty can be ntroduced f we use a planner that works wth nondetermnstc events, as shown n [3]. The reputable lterary scholar, Todorov, has convncngly argued [4] that detectve stores actually contan two stores: the story of the crme and the story of the nvestgaton. Accordng to hm, the characters of the story of the nvestgaton do not physcally act very much, they manly learn. In contrast, t so happened that n the Swords-and- Dragons genre, the frst to be treated n our Logtell nteractve plot-composton project [1], the plots were domnated by physcal act scenes, such as the kdnappng of a prncess amd volent combats between a vllanous dragon and the brave knghts who come to rescue her. Communcatve acts, ndspensable to extend our methods to other more subtle genres, were clearly mssng. Accordngly, a second objectve of the present paper s to provde an adequate set of operatons for accomplshng such acts, thus advancng a study whose prelmnary results were reported n [5]. The magnatve (and magnary) protagonst of our examples of detectve stores, Bertllon, fgures brefly n Sr Arthur Conan Doyle's The Hound of the Baskervlles 1, n probable alluson to a hstorc poneer n Anthropometry: 2 Dr. Mortmer: Recognzng, as I do, that you are the second hghest expert n Europe " Holmes: "Indeed, sr! May I nqure who has the honour to be the frst?" Dr. Mortmer: "To the man of precsely scentfc mnd the work of Monseur Bertllon must always appeal strongly." The remanng of the text s organzed as follows. Secton II descrbes our recently-developed non-physcal event-producng operatons. Secton III exposes our vew of the detectve stores 1 cf

2 40 SBC Journal on Interactve Systems, volume 5, number 2, 2014 genre, presents a summary of ts logc programmng specfcaton, and reports Bertllon s early feats. Two frendly user nterfaces, to nteractvely produce and dsplay the plots, are shown n Secton IV. Related work s brefly surveyed n Secton V, and Secton VI has the conclusons. A fuller descrpton of our logc-programmng model, whch adopts the Entty-Relatonshp model [6] to specfy facts, the STRIPS method [7] to defne events, and stuaton-objectve rules to motvate the characters' behavour, as well as detals about the system's mplementaton, can be found n a prevous techncal report. 3 II. EVENTS INVOLVING NON-PHYSICAL ACTS A. Informaton-gatherng events The nformaton-gatherng events [5] enable the varous characters to mentally apprehend the state of the world. Wthout such events, one would have to assume that the characters are omnscent. Here we shall recognze a sharp dstncton between the facts themselves and the sets of belefs [8][9] of each character about the facts that hold at the current state of the world, whch consttute, so to speak, ther respectve nternal states. Belefs can be rght or wrong, dependng on ther correspondng or not to the facts. Moreover, we have taken the opton that acqurng a belef does not cancel a prevous belef. As a consequence, we allow a character to smultaneously entertan more than one belef wth respect to the same fact, possbly wth a dfferent degree of confdence whch depends on the provenance of the belefs. We shall consder three types of nformaton-gatherng events, each type assocated wth a set of operatons: Communcaton events - operatons: ask, tell, agree, ask_event, tell_event, agree_event Percepton events - operatons: sense, watch Reasonng events - operatons: nfer, suppose. Operatons sense, ask, tell, agree, nfer, and suppose refer to belefs on facts, whereas watch, ask_event, tell_event, and agree_event refer to some acton event wtnessed by a character. The operatons are defned n terms of ther pre-condtons and post-condtons [7]. The precondtons are logcal expressons nvolvng affrmed or negated facts and belefs, whereas post-condtons denote the effect of the operaton n terms of belefs that are added or deleted to/from the current nternal states of the characters nvolved. However, the specfcaton of the operatons s delberately kept at a mnmum, to be complemented by separate condtoners that express the peculartes of the characters partcpatng n the stores. Wthn computer scence, communcaton between characters mmedately brngs to mnd the communcaton processes executed by software agents n mult-agent systems. In partcular, the Agent Communcaton Language (ACL) conssts of operatons smlarly defned by ther pre- and postcondtons [10]; for an earler more formal treatment, see for nstance [11]. Software agents dffer from fctonal characters 3 ftp://ftp.nf.puc-ro.br/pub/docs/techreports/12_08_barbosa.pdf (and, roncally, from human bengs n general) n that they are supposed to only transmt nformaton n whch they beleve, to agents that stll lack such nformaton and need t n order to play ther role n the executon of some practcal servce. In contrast, certan characters are prone to le, ether for ther beneft or even out of habt. In general they may gnore the conversatonal maxms prescrbed by phlosophers of language, such as [12]. The specfcaton of our tell(a,b,f) operaton does not even requre that A has any noton of the fact F to be transmtted to B. It s enough that both characters are at the same locaton L; f they are not, a current_place(a,l) sub-goal s recursvely actvated, whch may cause the dsplacement of the teller (character A) to L, where B currently s. And the only necessary effect of the operaton s that F has been told by A to B. Whether or not B wll beleve n F wll depend on the executon of the agree operaton, whch n turn depends on whether or not B trusts A. Another purpose served by tell(a,b,f) s to convey merely expressve speech acts [13]; the F parameter can then be any arbtrary sentence. The B parameter may reman unspecfed, n whch case A s addressng a general audence. The ask operaton s smlarly defned, and ts effect s just that A has asked F from B, who may respond or not. The fundamental character-dependent condtoners are establshed, respectvely, by separate wll_tell and wll_ask condtoners. Percepton s the faculty whereby people keep contact wth the world through ther fve senses (sght, hearng, touch, smell, and taste). At the present stage of our work we do not make such dstnctons, and merely consder a generc sense(c,f) operaton to apprehend any sort of fact F specfed n the statc schema as perceptble, wth a varant verson that makes provson for defectve sensng. For correct sensng of a postve or negatve fact F, F must be successfully tested. Dstorted sensng (for nstance, of certan colours by a daltonc subject) s accompaned by a sde-remark on the true fact. In any case, besdes the sensed clause, a belef clause s mmedately added, snce drect percepton does not depend on a thrd party. The watch(c,e) operaton allows a character C to wtness an event E, denoted as always n our system by some operaton defned n the dynamc schema. Operatons ask_event(c', C, E), tell_event(c, C', E), and agree_event(c,c',e) sgnfy, respectvely, that another character C' questons C about E, that C reports the event, and that C' effectvely agrees wth C about ts occurrence. As before, the defntons of these operatons are completed by condtoners, respectvely sense_rule and watch_rule clauses. For sense, t s requred that, to ascertan a postve or negatve fact F nvolvng a person or object currently at place L, a character C must be at L, ether orgnally or as the result of pursung current_place(c,l) as a sub-goal. For watch, the current_place requrements depend on the type of event beng watched, whch justfes ther beng left to the specal watch_rule clauses. For nstance, the operaton go(a,l1,l2) (requred to update current_place(c,l)facts), can be watched partly by persons at L1 (orgn) and at L2 (destnaton).

3 SBC Journal on Interactve Systems, volume 5, number 2, Both the agent of an event E and a character who watches the occurrence of E are, as expected, aware of the man effects of the event. Antcpatng what we shall treat n our detectve stores envronment, f A klls B, or f C watches A kllng B, then characters A and C wll beleve the facts klled(a,b) and dead(b,true). And when the agent or a wtness uses tell_event to nform another character C' and C' reacts wth an agree_event, C' wll also start belevng n the facts caused by the reported event. Deducton, nducton and abducton are complementary reasonng strateges. For deducton, f there s a rule A B and the antecedent A s known to hold, t s legtmate to nfer that the consequent B holds. In the case of nducton, the systematc occurrence of B whenever A occurs may justfy the adopton of rule A B. Abducton (cf. [14]) s a non-guaranteed but nevertheless useful resource n many uncertan stuatons: gven the rule A B, and knowng that B holds, one may suppose that A also holds. Ths s a type of reasonng habtually performed by medcal doctors to dagnose an llness from observed symptoms. The trouble s, of course, that t s often the case that more than one llness may provoke the same symptom,.e., there may exst other applcable rules A 1 B, A 2 B,..., A n B, suggestng dfferent justfcatons for B. Thus n abducton, wheren the mplcaton arrow s followed backward, one s led to formulate hypotheses rather than the frm conclusons ssung from deducton over determnstc rules. Our nfer and suppose operatons utlze, respectvely, deducton and abducton. Ther condtoners can be the same rules of nference (nf_rules) to be traversed forward n the former case or backward n the latter. In our mplementaton of the nfer operaton, gven a rule P=>F accepted by character A, the antecedent P furnshes the belefs to be tested as precondton, whereas A s belef n F wll be acqured as the added effect (another addton beng an nferred clause) upon a successful evaluaton of P. In contrast, n the case of the suppose operaton, the belef n the consequent wll just motvate the addton of a supposed clause n a fact present n the logcal expresson of the antecedent. We must stress that the nference rules adopted by the characters n our story context do not pretend to be scentfcally correct. Often orgnatng from popular tradtons, they may lead to farfetched or absurd belefs. Indeed wth a nave nference rule, establshng that the colour red would be mstakenly perceved as green by daltonans, and wth the nterplay of credulous persons wth both honest and lyng nformers, a varety of plans may result, some qute unexpected. Posng to our planner the goal of fndng someone who would (wrongly) beleve that Maran's har was not red:?- plans((beleves(x,har_colour('maran',c)), not (C == red)),p), narrate(p). the planner frst gave us two plans where X was nstantated wth Peter, declared to be daltonc and always ready to accept whatever anyone would tell hm. In the shortest plan, he would smply look at red-hared Maran and sense that she had green har. In the second, Jane, a notorous lar, would come from Manchester to London to tell hm that Maran was blond. But the thrd plan really came as a surprse: start=>sense(maran,daltonc(peter,true))=>tell(john,peter,har_colour(maran,red))=>agree(peter,john,ha r_colour(maran,red))=>ask(maran,peter,har_colour (Maran,_))=>tell(Peter,Maran,har_colour(Maran,re d))=>nfer(maran,har_colour(maran,green)) whch the narrate command translated n template-drven natural language as follows: Maran senses that Peter s daltonc. John tells Peter: "- Maran has red har". Peter agrees wth John. Maran asks Peter: "- What s the colour of my har?". Peter tells Maran: "- Your har s red". Maran nfers that she has green har. In the past, dealng wth the Swords-and-Dragons genre, descrbed n [1], we were faced wth an even more ntrgung result, when challengng the planner wth what we thought was an mpossble goal. In words, we requred that the man hero should fal to kll the dragon, addng that he would treat n a dscourteous manner the magcan who would endow hm wth the necessary strength. About the magcan, we must add that, accordng to our specfcaton, he would penalze dscourtesy wth the opposte effect, decreasng to a crtcal level the postulant's resstance. The planner dd not take much tme to come up wth a soluton: an even weaker knght went to fght the dragon. Though not succeedng, as one could expect, he managed to tre the dragon, thereby reducng ts strength. Then came the hero, now strong enough to kll the beast wthout extra power. And what about the magcan? Only after kllng the dragon, the hero went to the forest where the mghty enchanter lved, and proceeded to mstreat hm... B. Drectve and commssve events In order to acheve a desred goal, a character C may need to resort to another character C' to perform an acton that C s unable or unwllng to execute personally. Three operatons were suppled to meet ths requrement: request, comply, and refuse. Lngusts [13] classfy such communcatve acts n the drectve (the frst one) and n the commssve categores (the two last ones). As before, the specfcaton of pre-condtons and postcondtons s complemented by condtoner clauses, named respectvely wll_request, wll_comply and wll_refuse. It s normal to specfy that C' always comples to the requests of C f an obeys(c',c) relatonshp has been declared to bnd them. On the other hand, complance may be subjected to a recprocal request: C' would, so to speak, negotate wth C, mposng a task as payment or compensaton for the servce to be rendered to C. Moreover, even f a character has compled to perform an acton, t does not necessarly follow that the promse wll be fulflled, whch s n consonance wth the exstence of characters who shamelessly le. The request operaton can be ether drected to a specfc character or left open, so that we acheve the generalty afforded by the cfp (call-for-partcpaton) operaton of ACL [10].

4 42 SBC Journal on Interactve Systems, volume 5, number 2, 2014 C. Lbrary-consultng events As a dual process to plan-generaton, plan-recognton s a no less nvaluable resource for the composton of story plots. In prncple, f we let the plan-generator run for an ndefnte amount of tme, t should produce all plots consstent wth the gven specfcaton, even those unworthy of our attenton. To offer an alternatve, that may be hander n some stuatons, we provde a collecton of story patterns extracted from preexstng narratves of dverse provenance. Such collecton, resdng n a convenently ndexed lbrary [15][16][17], could then contrbute to create new plots by adaptaton, or to match a few observed events aganst typcal story patterns, thereby allowng to predct what the agents are tryng to accomplsh. Our lbrary tems obey the format user/man-agent/ conclusve-message/goals/plan/complementary-test. For our current purposes, the plan, the test and the message are of specal nterest. Our lbrary-consultng operatons, whch are also ncorporated n the plots produced by the planner, are named collect, recognze, and try. The frst smply assembles n a lst the events that have been drectly watched by or related to a character. The recognze operaton checks whether all events n the lst of observatons correspond to events n the plan component of a lbrary tem. If the patternmatchng succeeds, the logcal expresson n the test s passed to the try operaton, () to exclude trval cases of successful matches and () to trgger addtonal events, such as recommendatons for the detectve to gather further nformaton. Such events are then appended, thus becomng part of the generated plot. Fnally, f both the pattern-matchng and the subsequent test succeed, the message s composed and becomes avalable (n partcular to the doman-orented expose operaton to be ntroduced n part B of the next secton, whereby the detectve communcates hs conclusons). III. THE GENRE OF DETECTIVE STORIES A. Detectves n acton One thng we do not propose to do s to model the acton of real-lfe detectves, whose work s grounded today n the hghly sophstcated resources of forensc scence. Our approach to the genre of detectve stores s based on the processes adopted by some llustrous fctonal detectves. Curously varous suggestons from such experts are n harmony wth a major contrbuton of Semotcs, namely the characterzaton of the so-called four master tropes, proposed n the past by, among others, Petrus Ramus and Gambattsta Vco, and revved n our tmes by Kenneth Burke [18]. In prevous work we assocated these tropes metonymy, metaphor, synecdoche, and rony wth, respectvely, four relatons between events, whch we have denomnated [19] syntagmatc, paradgmatc, meronymc, and antthetc. They have been declared to consttute "a system, ndeed the system, by whch the mnd comes to grasp the world conceptually n language" [20]. Accordng to [21], metonyms are based on varous ndexcal relatonshps between concepts, notably the substtuton of effect for cause, and convey an dea of contguty. Borrowng from [22], we requre the presence of syntagmatc relatons between events, to justfy ther beng meanngfully placed n sequence. Indeed a detectve must frst of all see the events under nvestgaton as a coherent causeand-effect sequence, wheren each event creates the condtons for what comes next. Dupn's method s of ths sort [23]: At such tmes I could not help remarkng and admrng (although from hs rch dealty I had been prepared to expect t) a pecular analytc ablty 4 n Dupn. He seemed, too, to take an eager delght n ts exercse f not exactly n ts dsplay and dd not hestate to confess the pleasure thus derved. The defnton of events va the pre- / post-condtons of operatons and the composton of plots by a backwardchanng planner s the man devce we use to guarantee consstency along the syntagmatc axs. Our stuatonobjectve rules also play an mportant role, functonng as trggers for future actons of the agents nvolved. Specfcally for the doman of detectve stores, we concentrate on motvaton aspects. The paradgmatc relatons, nspred on metaphor [24], arse from smlartes and analoges. Story patterns tend to repeat themselves, as Hercule Porot so well realzed when reflectng on Norton's skll to nduce several other people to commt a crme n hs stead [25]: It was amazng. But t was not new. There were parallels. And here comes n the frst of the "clues" I left you. The play of Othello. For there, magnfcently delneated, we have the orgnal of X. Iago s the perfect murderer. The deaths of Desdemona, of Casso ndeed of Othello hmself are all Iago's crmes, planned by hm, carred out by hm. As mentoned n secton II, part C, our specfcatons nclude lbrares of story patterns, to be accessed by lbraryconsultng events, thus allowng detectves to revew prevous cases, classc or not, that seem to ncorporate some famlar motf. In [26], where sx types of part-of lnks are dstngushed, one reads: "We wll refer to relatonshps that can be expressed wth the term 'part' n the above frames as 'meronymc' relatons after the Greek 'meros' for part". Gong down to detals, such as fndng how many tmes the ash of a cgar has fallen on the sol, s one of Sherlock Holmes precautons [27]: Before turnng to those moral and mental aspects of the matter whch present the greatest dffcultes, let the enqurer begn by masterng more elementary problems. Let hm, on meetng a fellow-mortal, learn at a glance to dstngush the hstory of the man, and the trade or professon to whch he belongs. Puerle as such an exercse may seem, t sharpens the facultes of observaton, and teaches one where to look and what to look for. By a man's fnger nals, by hs coat-sleeve, by hs boot, by hs trouser 4 The emphass s ours

5 SBC Journal on Interactve Systems, volume 5, number 2, knees, by the callostes of hs forefnger and thumb, by hs expresson, by hs shrt cuffs by each of these thngs a man's callng s planly revealed. That all unted should fal to enlghten the competent enqurer n any case s almost nconcevable. Noteworthy detals thus nclude fngerprnts, footprnts, lkely and unlkely weapons, the fact that someone s carryng a jewel, etc. We have dealt elsewhere [19], though not here, wth an even more sgnfcant aspect of whole-part decomposton, namely the descrpton of events at the level of more basc actons. For example, a homcde may comprse the obtenton of a lethal poson, the act of pourng t n a glass of wne, etc., etc. Antthetc relatons express negaton and opposton, such as the apparently rreducble dfference between good and evl and, consequently, between the hero and the vllan. And yet roncally shftng from one extreme to ts contrary may be necessary for understandng an opponent. Father Brown, a catholc prest, a man of mpeccable morals, thus explans hs performance as a detectve [28]: I had planned out each of the crmes very carefully," went on Father Brown, "I had thought out exactly how a thng lke that could be done, and n what style or state of mnd a man could really do t. And when I was qute sure that I felt exactly lke the murderer myself, of course I knew who he was. Qute approprately, we must remember that the learned chroncler of Father Brown's adventures was consdered a master of paradox. But dramatc rony [29] can perhaps be ponted out as the most characterstc ngredent of detectve stores. The character who looks more nnocent-lookng s n many cases found to be the sought-for crmnal. And frequently the detectve s compelled to change a lne of nvestgaton because t s revealed that thngs were "another way round". What makes a story nterestng s almost always the culprt's skll as a decever, nducng false belefs that untl the fnal showdown appear to be true. Among the early cases of Bertllon to be reported n part D, the two last ones would seem to have a touch of rony. Let us consder one more pont about the habtual reasonng practces of a detectve. Lke medcal doctors, they often proceed by abducton, on whch we based our suppose operaton. Sherlock Holmes once sad to hs frend Dr. Watson, as the good doctor narrates n hs memors [27]: "... In solvng a problem of ths sort, the grand thng s to be able to reason backwards. That s a very useful accomplshment, and a very easy one, but people do not practse t much. In the every-day affars of lfe t s more useful to reason forwards, and so the other comes to be neglected. There are ffty who can reason synthetcally for one who can reason analytcally." "I confess," sad I, "that I do not qute follow you." "I hardly expected that you would. Let me see f I can make t clearer. Most people, f you descrbe a tran of events to them, wll tell you what the result would be. They can put those events together n ther mnds, and argue from them that somethng wll come to pass. There are few people, however, who, f you told them a result, would be able to evolve from ther own nner conscousness what the steps were whch led up to that result. Ths power s what I mean when I talk of reasonng backwards, or analytcally." To fnsh ths secton, let us recall how Mrs. Aradne Olver's ntuton surprsngly seemed, whle apprasng Porot's methods, to antcpate what Bertllon would propose to do wth the beneft of the more advanced technology of our 21 st century [30]: "Do you know what you sound lke?" sad Mrs. Olver. "A computer. You know. You're programmng yourself. That's what they call t, sn't t? I mean you're feedng all these thngs nto yourself all day and then you're gong to see what comes out." "It s certanly an dea you have there," sad Porot, wth some nterest. "Yes, yes, I play the part of the computer. One feeds n the nformaton. " "And supposng you come up wth all the wrong answers?" sad Mrs. Olver. "That would be mpossble," sad Hercule Porot. "Computers do not do that sort of a thng." "They're not supposed to," sad Mrs. Olver, "but you'd be surprsed at the thngs that happen sometmes. My last electrc lght bll, for nstance. I know there's a proverb whch says To err s human, but a human error s nothng to what a computer can do f t tres." B. Events n our detectve stores In our subclass of the detectve stores genre, two events correspond to crmes: kll and steal. Event attack s also an aggressve acton, whose agent can n prncple be any of the characters. For the detectve's provsonal or fnal conclusons an expose event s provded. The specfcaton of operaton kll, the major focus of all nvestgatons reported n part D, s shown below. The pre-condton combnes some motvaton clause wth the general requrements that the vctm should not already be dead and that the kller must be at the same locaton as the vctm. Each fact F mentoned n the Crc parameter (the crcumstances that motvate the crme) of the motvaton clause s converted nto beleves(x,f), where X s the wouldbe crmnal. operaton(kll(x,y,m)). added(klled(x,y),kll(x,y,m)). added(motve(x,[kll(x,y),m]),kll(x,y,m)). added(dead(y,true),kll(x,y,m)). precond(kll(x,y,m),p) :- motvaton(x,[kll(x,y,m),crc]), prep_mot(x,crc,crc1), appc((current_place(x,l), /current_place(y,l), /(not dead(y,true))), Crc1,P). A few typcal motvaton clauses follow: motvaton(a,[kll(a,b,greed),(owns(b,o), not (A = B), carres_object(a,o))]). motvaton(a,[kll(a,b,jealousy),(loves(a,b), loves(b,c),gender(a,m),gender(c,m),

6 44 SBC Journal on Interactve Systems, volume 5, number 2, 2014 not (A = B), not (A = C))]). motvaton(a,[kll(a,b,request),(obeys(a,c), not (B = C), compled(a,[c,kll(a,b,request)]))]). motvaton(a,[kll(a,b,vengeance), (loves(a,c), not (A = B), not (A = C), klled(b,c), not (B = C))]). motvaton(a,[kll(a,b,'self-defense'), (attacked(b,a), not (A = B))]) :- A = 'Maran'. motvaton(a,[kll(a,a,lovesckness),(loves(a,b), not loves(b,a), not (A = B))]). A detectve s supposed not only to fnd the dentty of the kller but also the motve of the crme. An economcal way to formulate an approprate nference clause, to be appled by detectve D, s, n words: D wll nfer that, f D beleves that X klled Y and that the crcumstances descrbed n Crc currently held, then the motve M that guded X was that ndcated n the motvaton clause correspondng to those same crcumstances. Two clauses were used to express that, the second clause servng to ntroduce a bas: we antcpate that, n the case of sucde, our detectve wll always assume a crss of lovesckness as explanaton for the kller/vctm s desperate acton: nf_rule(d,(klled(x,y),crc) => motve(x,[kll(x,y),m])) :- motvaton(x,[kll(x,y,m),crc]), not (X == Y). nf_rule(d,klled(x,x) => motve(x,[kll(x,x),lovesckness])). As sad before, detectve stores make ample use of all the communcatve events of secton II. Among the condtoners that provde the necessary flexblty to the operatons, those assocated wth the drectve and commssve events deserve specal attenton. In the context of detectve stores, such events moblze the relatonshp between nstgators and ther accomplces. The followng wll_comply condtoners convey a possble arrangement: a character C1 who obeys C2 wll always comply wth whatever C2 may request; on the other hand, f the obedent C1 requests that the domnatng C2 shall kll some person C3, C2 wll comply only f C1 n turn comples to also get nvolved n the crmnal aggresson aganst C3, specfcally by stealng an object owned by the vctm (such trcky negotaton occurs n case 6 of part D): wll_comply(c1,c2,act,obeys(c1,c2)). wll_comply(c2,c1,kll(c2,c3,request), (obeys(c1,c2), owns(c3,o), compled(c1,[c2, steal(c1,o,c3)]))). C. Enter Bertllon the context of hs mnworld The dramats personae fgurng n our example are: Bertllon - a prvate French detectve ntatng hs career n England Maran - lovely red-hared young lady, stll unmarred Robn - Maran s sutor Patrck - another sutor of Maran, a notorous lecher Jane - former actress servng as Patrck s accomplce Cogsworth - Brtsh butler, head of Maran s household At the ntal state, all characters are n London, except Jane, who s n Manchester. Cogsworth, Maran s loyal butler, s destned to partcpate as Bertllon s man wtness. He s ever dsposed to testfy and to volunteer nformaton, always consstent wth what he beleves to be true. Maran s equally sncere, but a shade too credulous. She accepts whatever s told by Jane, who happens to be a compulsve lar. Maran often plays the role of the vctm. Patrck s tempted to kll her, ether mpelled by jealousy or because he covets a precous jewel that she mprudently wears attached to a necklace. She must also beware of Jane, obsessvely averse to red-hared persons and an obedent accessory to Patrck s machnatons. On the brght sde, she has nothng to fear from her butler, and enjoys a recprocal love relatonshp wth Robn. Cases 1 through 5 are rather straghtforward. Manly relyng on Cogsworth s testmony but also on hs (un)far knowledge of the meta data Bertllon s able to nfer who s the culprt and hs or her motvaton. Cases 6 and 7 are somewhat more nvolved, compellng the detectve to resort to our lbrary of crme patterns. The lbrary tem for case 7 s: lb([... (U/A/[A,' deceved ',B,', possbly wth crmnal ntent']/(loves(b,c), told(a,[b, not loves(c,b)]), beleves(b, not loves(c,b)), told(a,[b, loves(c,d)]), beleves(b, loves(c,d)), klled(b,c),klled(b,b))/ (start => tell(a,b,not loves(c,b)) => agree(b,a, not loves(c,b)) => kll(b,c,jealousy) => kll(b,b,m))/(/told(a,[b,not loves(c,b)]), trusts(u,x), asked(u,[x,loves(c,b)]), told(x,[u,loves(c,b)]), agreed(u,[x,loves(c,b)]))),...]). D. Hs seven cases For each case, we shall provde a bref synopss. To save space, we gve only for the frst case a callng sequence to the planner able to generate the plot, and the natural language textual renderng of the plot. Such texts result from the applcaton of stll rather crude templates [31], whch we ntend to mprove at a later phase of the project. case 1: In a ft of passon. Patrck klls Maran, unaware of the presence of the butler, who s later n a poston to communcate the event to Bertllon, together wth all nformaton needed to establsh that the murderer s motve was jealousy. callng sequence: ex1 :- plans(( motve('patrck',[kll('patrck','maran'), jealousy]), watched('cogsworth',kll('patrck','maran',m)), related('cogsworth',['bertllon', kll('patrck','maran',m)]), agreed_op('bertllon',['cogsworth', kll('patrck','maran',m)]), agreed('bertllon',['cogsworth', loves('patrck','maran')]), agreed('bertllon',['cogsworth',

7 SBC Journal on Interactve Systems, volume 5, number 2, loves('maran','robn')]), nferred('bertllon', motve(s,[kll(s,'maran'),m])), exposed('bertllon',[s,'maran',jealousy,nl]) ),Plan),narrate(Plan), nl, nl,!. plot n template-drven natural language: Patrck senses that Maran loves Robn. Patrck klls Maran. Cogsworth watches the event: Patrck klls Maran. Cogsworth relates to Bertllon the event: Patrck klls Maran. Bertllon agrees wth Cogsworth about the event. Cogsworth tells Bertllon: Patrck loves Maran. Bertllon agrees wth Cogsworth. Cogsworth tells Bertllon: Maran loves Robn. Bertllon agrees wth Cogsworth. Bertllon assumes that Patrck, n the event Patrck klls Maran, was motvated by jealousy. Bertllon says: The suspect s Patrck and the motve s jealousy. case 2: All that gltters. Ths tme Patrck commts two crmes: murder and theft. The frst s once agan watched by the butler, who later notces that the culprt carres the object of the theft, thus characterzng greed as the prmary motvaton. case 3: Murder by proxy. Patrck procures Maran s death by orderng Jane to do the kllng. The butler watches ther conversaton and Jane s fatal act. Learnng of both scenes from Cogsworth, Bertllon establshes Jane s drect nvolvement and Patrck s role as nstgator, not botherng however to dsclose hs motve (jealousy). case 4: The reluctant vctm. But our vctm s not necessarly so helpless! Jane, on her own ntatve, repelled as she s at the sght of red-hared Maran, attacks her to be promptly klled n reacton. The butler watches the aggresson and the counter-aggresson events, whch, reported to Bertllon, result n a verdct of self-defense. case 5: Avengng fury. Patrck klls Maran moved by jealousy. Two persons watch the murder: her butler and her lover. Cogsworth lmts hmself to gve testmony, but Robn s reacton s more effectve: he klls the assassn. To many a body of jurors, one mght surmse, vengeance n the aftermath of a henous murder should seem admssble as extenuatng crcumstance. case 6: Framed! Patrck s loose morals lead hm to arm a trap for the submssve Jane. Preparng to kll Maran, pressed as before by hs jealous mpulses, he takes advantage of Jane s own nclnatons. To extermnate the detested red-hared young lady, Jane requests the ad of her superor. As a condton to comply, Patrck requres that she should also perform some aggressve act, whle makng a proft; namely, she must steal Maran s rch jewel. The butler msses Jane s request to Patrck, but watches Patrck mposng the theft of the jewel as a condton for complyng, and also the theft tself. However, he reports to Bertllon only the latter event whch n no way ncrmnates Patrck. Percevng that Maran s dead, Bertllon, lke so many detectves n popular fcton, bulds the hypothess (an nstance of abductve reasonng from the only rule he knows to explan death: klled(x,y) => dead(y,true)) that someone klled her. And, at a loss for anythng else, he proceeds from the only clue avalable,.e. the theft reported by Cogsworth, to consult hs lbrary of crme patterns. A match occurs wth a pattern nvolvng requested theft by one person as cover-up for the more drastc act of the nstgator. The lbrary tem recommends checkng whether such a request occurred, leadng Bertllon to queston Cogsworth, who responds relatng the event that he, at frst, had faled to report. Once, for a change, Bertllon has not enough data for a frm nference, but he advances the possblty that Patrck murdered Maran and sought to cast all suspcon upon hs accomplce, Jane. case 7: Iago syndrome. Though consentng to the whms of Patrck, Jane surpasses hm by her far rcher magnaton. Patrck s not even mentoned n ths case, the hardest one untl now n Bertllon s career. Jane les to Maran, convncng her that Robn does not love her. In consequence, Maran s overcome by lovesckness and klls herself. The ll-ntentoned conversaton and ts fatal outcome are watched by Cogsworth, and ths tme he at once reports both scenes to Bertllon. The verdct of sucde s nescapable and Bertllon clearly states t. Feelng, however, that Jane s talkng to the vctm may have further mplcatons, he agan resorts to the lbrary. What he fnds s the same pattern that took hs Belgan colleague, Hercule Porot, to dentfy Norton as the wonted X n hs last case [25]. The located lbrary tem contans a recommendaton, leadng hm to check wth the well-nformed Cogsworth: does Robn love Maran, contrary to what Jane had proclamed? The butler s affrmatve reply s an ndcaton that Jane delberately nduced Maran to commt sucde. Bertllon denounces Jane, despte hs convcton that no Brtsh court of justce would condemn her. IV. INTERACTIVE COMPOSITION AND DRAMATIZATION Hopefully the cases narrated n the prevous secton are suffcent to gve an dea of what can arse from the current specfcaton of detectve stores, and hence offer some ndcaton of how useful s the package of communcatve speech acts descrbed before. However, generatng an entre plot va a sngle call to the plan-generator algorthm s not satsfactory from a dgtal entertanment vewpont. An envronment to run plot composton nteractvely s essental. In the course of our Logtell project, we developed several nterfaces to allow the user to create plots nteractng wth the plan-generator, whch also permt vsualzaton, dsplayng the generated plots as frame pcture sequences n storyboard style [19], or n comc book format [32], or by employng more expressve anmaton [1] and vdeo-based [33] technques. We shall now descrbe how nteractve composton can be acheved, on ordnary computers and on tablets, wth the support of the two smpler forms of dramatzaton mentoned above. A. Wth PlotBoard The flow of control of our PlotBoard tool [19] s shown below (Fgure 1). The tool's plan-based desgn, as n the other Logtell products, keeps playng a fundamental role, snce t ncorporates a knowledge of the doman that a casual user may not possess, but now the plan-generaton algorthm serves n a secondary helper's capacty. Under the user's command, the tool generates the plot n a stepwse fashon, alternatng between the user mode and the planner mode.

8 46 SBC Journal on Interactve Systems, volume 5, number 2, 2014 start user u: fnsh u: alternatve compose planner u: OK plan step u: show effects plot u: fnsh u: valdate show use gven plot submt u: accept end use plot from lbrary adapt Fg. 1. Flow of control of PlotBoard. When the planner mode s on, the stuaton-objectve rules are actvated to fnd the short-range goals that can, at the current state, be pursued by the planner. Pckng one of these goals, the planner then produces an approprate plan and dsplays t to the user, who may ssue an ok or ask for an alternatve plan (for the same or for another smultaneously actve goal). After the chosen plan step s executed, the user s asked whether the planner may contnue, n whch case the stuaton-objectve rules are agan actvated for the next step. But the user can prefer to shft to user mode, wheren several types of nterventon are announced n a menu, such as ndcatng the goal to be acheved next by the planner or even a specfc operaton to be executed. And, after sgnalng fnsh, the user can stll decde to perform one or more adaptatons of several knds over the events, such as nsertng, deletng, replacng, reorderng, summarzng, detalng, etc. The stuaton-objectve rules adopted n our tral run are lsted below, prefxed wth numbers for easy reference: (1) st_obj('patrck', (loves('patrck',y),not dead(y,true), not loves(y,'patrck'), loves(y,z), not ('Patrck' = Z)), (motve('patrck',[kll('patrck',y),jealousy]))). (2) st_obj('jane', (owns(y,o),not dead(y,true),carres_object(y,o)), (sensed('jane',owns(y,o)),carres_object('jane',o), motve('jane',[kll('jane',y),greed]))). (3) st_obj('jane', (har_colour(v,red),not dead(v,true),loves(h,v), loves(v,h)),(told('jane',[v,not loves(h,v)]), watched('cogsworth',tell('jane',v,not loves(h,v))), agreed(v,['jane',not loves(h,v)]))). (4) st_obj('maran', (not dead('maran',true), loves('maran',h), beleves('maran',not loves(h,'maran'))), (loves(m,'maran'),not (M = H), told('maran',[m,'let us meet someday!']))). (5) st_obj('maran', (not dead('maran',true), loves('maran',h), beleves('maran',not loves(h,'maran'))), (klled('maran','maran'), watched('cogsworth',kll('maran','maran',_)), beleves('cogsworth',klled('maran','maran')))). (6) st_obj('cogsworth', (beleves('cogsworth',klled(v,v))), (related('cogsworth',['bertllon', tell(s,v,not loves(h,v))]),agreed_op('bertllon', ['Cogsworth', tell(s,v,not loves(h,v))]), related('cogsworth',['bertllon',kll(v,v,m)]), agreed_op('bertllon',['cogsworth',kll(v,v,m)]))). (7) st_obj('bertllon', (beleves('bertllon',klled(v,v))), (nferred('bertllon',motve(v,[kll(s,v),m])), exposed('bertllon',[s,s,m,nl]))). (8) st_obj('bertllon', (beleves('bertllon',klled(v,v)), exposed('bertllon',[s,s,m,nl])), (obs('bertllon',obs),recognzed('bertllon', 'Bertllon'/Ag/Cr_type/Goals/Pl_lb/Q), tred('bertllon',q), exposed('bertllon',[s,v,lb,cr_type]))). At the ntal state, rules (1), (2) and (3) offer alternatve possbltes to the user's choce. If the goals of ether (1) or (2) are pursued, Maran s murdered (by Patrck or by Jane, respectvely). If (3) s preferred, Jane comes from Manchester and tells a le to Maran n the butler's presence, concdng wth the frst events of case 7, as descrbed n the prevous secton. The entre plot of case 7 wll ndeed be generated f one chooses, at each subsequent step, the alternatves offered by rules (5) through (8). But the story does not have to end badly. After applyng (3), rules (4) and (5) become actve. Wth (5) Maran yelds to depresson, but (4) allows her to recover and teach the (allegedly) dsloyal Robn a lesson, by gvng a chance to hs rval, the ll-reputed Patrck. When we started the tral run shown n fgure 2, we took the planner mode and, from the alternatve event sequences generated, chose the one produced by rule (3), and then shfted to the user mode. In a drect nterventon, we added an attack(maran,jane) event, whereby Maran somehow reacted to Jane's provocaton. But, returnng next to the planner mode, we chose (5) so that the sucde scene ensued, agan wtnessed by the horrfed Cogsworth. At ths pont, we ndcated that the generaton phase was fnshed. Asked by the tool whether the obtaned plot should be accepted, we chose nstead the adapt opton, and removed the ffth event (n whch

9 SBC Journal on Interactve Systems, volume 5, number 2, Maran expressed her belef n Jane's false asserton). We then selected the show opton from the menu, thus causng the plot to be dsplayed va the Prolog/Java nterface. Fg. 2. Interactve composton wth PlotBoard. B. Wth Comc Books We now turn to an nteractve storytellng system capable of representng the generated detectve stores n the format of comc books on tablet computers, where users are able to nteract wth certan objects that can affect the unfoldng stores (Fgure 3). Comcs are a classcal form of vsual storytellng that combnes mages and text, and s capable of evokng strong emotonal reactons from readers, creatng dentfcaton, and conveyng a story n a very appealng manner [34]. The system s based on a clent-server archtecture: the server hosts the planner responsble for generatng stores, and the clent contans the vsualzaton and nteracton nterface that presents the narratve n the format of comc books. The process of representng the generated stores through comcs conssts of three man phases: plot structurng, panel defnton, and panel compostng. In the plot structurng phase, the plot s organzed nto two storylnes: crme and nvestgaton. The story of the crme s presented frst, but wthout revealng the crmnal. More detals about the crme are revealed durng the nvestgaton, wth the detectve s nterventon. The panel defnton phase comprses the process of dynamcally assgnng the story events to ther correspondng panels, computng the sze requred for each panel, and defnng the layout of each page. And the panel compostng process conssts of gatherng all the vsual elements together to form the fnal mage of the panel. Panels are used to present a sngle moment frozen n tme. Lettng refer to tme, a panel P represents a dscrete nstant t (Fgure 3). Besdes the assocaton wth tme, the specfcaton of panel P comprses a specfc locaton L and a set of events E : P ={L,E }, E ={e,1, e,2,, e,j,, e,ne }. An event e s an nstance of a plannng operaton (e.g. kll('patrck', 'Maran','jealousy')). Events are always sequental n tme (.e. the story planner does not generate parallel events), but ths sequence s compressed n the dscrete nstant of tme t represented by a panel P. R k D d Panels gutter P h k H P 1 P 2 P dscrete t 1 t 2 t tme Fg. 3. Elements of a comc book page of sze D x H, where the panel P s wthn row R k and has a sze of d x h k. We establsh the followng rules to decde whether or not a new event e n can be grouped wth ts precedng event e p n a panel P, wthout breakng the contnuty of tme and space: 1. If e n and e p are both speeches of the same character, or dfferent characters that are at the same place, and the number of speeches already added to the panel P s smaller than α (maxmum number of speeches supported by a sngle panel), then the event e n can be assgned to the panel P.

10 48 SBC Journal on Interactve Systems, volume 5, number 2, If e n s a speech and e p contans an acton performed by the speakng character, then the event e n can be assgned to the panel P. 3. If e p and e n are the same event (whch s not a speech) performed by two dfferent characters at the same place, then the event e n can be assgned to the panel P. 4. Otherwse, e n s assgned to a new panel. Once the events have been assgned to ther respectve panels, the system can start creatng the pages that wll support the panels. A page s composed of a sequence of panels wth varyng sze and locaton, whch present the story events to the reader. The sze of a panel s generally proportonal to the amount of narratve content presented, and ts poston s relatve to the chronologcal order of the events. In order to dynamcally calculate the sze of the panels, we propose a method to estmate the mportance of a panel based on s assocated wth the events and the locaton where the events take place. Each class of event (e.g. go, kll, tell) and the locatons where the events can happen are assocated wth a based on ther mportance to the narratve. For example, a go event (a character goes from one place to another) may have less mportance to the narratve than a kll event (a character klls someone); also, some places are more mportant than others. These assgnments can be made by a sngle numercal value or a condtonal expresson (e.g. a go event may have ts ncreased f certan specfc events occur at tme t represented by panel P ). Therefore the s are calculated by a functon that depends on P. Weghts are also calculated for each row and, fnally, for the whole page. The followng equatons calculate the s of a panel P, a row R k, and a page F j : P R F L k j NK j 1 NR k 1 ( P ) P NE j, j 1 e j, j ( P ) P R Rk, NR where L ( P ) s a functon that returns the of the locaton L P ; e ( P ) s a functon that returns the of the event, j j E, E P ; NE s the number of events n, e the panel P ; NK s the number of panels n a row R k ; NR s the number of rows n a page; β s the maxmum value allowed for a row; and γ s the maxmum number of rows allowed n a page. The algorthm that calculates the sze of the panels and the layout of the pages starts by teratng through the panels and assgnng them to a page and a row accordng to ther chronologcal order. When the of a row (sum of the row panels s) reaches β (maxmum allowed to a row), the panels begn to be assgned to next row. When the number of rows reaches the maxmum number of rows per k page (γ), the panels begn to be assgned to the frst row of the next page. The algorthm ends when all panels are assgned to a page and a row. We must notce that the parameters β and γ determne the general aspect of the comc book page. In our prototype we assume β =6, and γ =3. The actual sze of each panel s calculated accordng to ts and poston n the page. The wdth d of a panel P n a row R k and the heght h of a row R k n a page are gven by: P d ( P, Rk ) D gutter P R R, k Rk hk ( Rk, F j ) Fj H gutter where D s the horzontal sze of the page and H s the vertcal sze of the page n pxels, and gutter s the space between panels. Panels can be composed of four types of objects: background layers, characters, nteractve objects, and text balloons. Background layers are a representaton of the envronment where the events occur. Every avalable locaton of the story s assocated wth a set of statc or dynamc mage layers that are used to create the scenaros of the story. Characters are composed of a set of behavours representng the actons they can perform durng the story. Each behavour comprses a set of statc mages representng the acton from dfferent angles. Durng the compostng process, the behavour s selected accordng to the acton performed by the characters, and ts poston and angle are defned by wayponts postoned n the scenaros. Interactve objects are composed of two mages, shown before and after the user nteracton wth them. Durng the compostng process, the objects are added to the panels as part of the scenaros. Text balloons are dynamcally generated and nserted n the panels respectng the followng rules: 1. B B n : Balloons should not overlap each other; 2. B C : Every balloon B should not overlap any of n the characters C n ; 3. B O : Every balloon B should not overlap any of n the nteractve objects O n ; 4. B must be placed accordng to ts chronologcal and readng order. Once a plot s generated by the plannng algorthm, the correspondng panels representng the story events are created, and users can read the story as a tradtonal comc book. Moreover, some scenaros nclude nteractve objects that can be actvated by tappng on them. Whenever such objects are present, the logcal context of the story can be modfed at that pont of the narratve, accordng to the effects assocated wth the actvated object. Reactng to the user s touch, the system requests a new plot from the plan-generator algorthm to create an alternatve story consstent wth the changes caused by the user nteracton. As a result, the effects of the user s nterventon are propagated to the next story events, and the k

11 SBC Journal on Interactve Systems, volume 5, number 2, comc panels are updated to reflect the new storylne. The user nteracton, whereby Bertllon s case 1 (murder motvated by jealousy) was converted nto case 2 (murder motvated by greed) s llustrated n Fgure 4b, whch shows the user s fnger pontng to Jane s precous jewel. (a) (b) (c) Fg. 4. The nteractve comc book: (a) a sample page; (b) the user touchng an nteractve object; (c) a page contanng the modfed events resultng from the user nterventon. V. RELATED WORK There have been a number of story generaton systems whch deal wth plot compostng usng a varety of dfferent methods. One of the earlest examples of an automated story generaton system s Tale-Spn [35][18], whch demonstrated the effectveness of usng plannng algorthms for the generaton of coherent characters and plots. Other examples nclude Unverse [36][36], whch added restrctons and authoral goals to the plannng algorthms n order to provde more control over the generated stores, and Mnstrel [37], whch proposed the reuse of parts of a story or an event lbrary n order to generate new plots. Mystery and, more specfcally, detectve stores do not consttute an unexplored feld n dgtal entertanment. In fact, plot-drven detectve stores (often called "whodunts",.e. "who done [dd] t?") have been extensvely used as the narratve envronment for many game-orented systems. One of the earlest examples of an automated storytellng system s the Automatc Novel Wrter [38][39], whch was programmed n FORTRAN and was capable of producng farly long murder mystery stores. The system reles on a smulaton model where the behavor of ndvdual characters and events s defned by probablstc rules formulated by the authors, whch progressvely change the state of the world. The system receves a descrpton of the world n whch the story s to take place as nput, together wth a descrpton of character trats that defne the murderer and the vctm. The motves arse as a functon of the events durng the course of the story. Another early example s the laserdsc game Murder, Anyone?, 5 played by two teams, whose goal s to guess the character Derrck Reardon s murderer, the motve and the method [39]. Tea for Three s a whodunt (nspred by Infocom s Deadlne [40]) where the user plays the role of a detectve who has to fgure out (by seekng physcal clues and talkng to characters) f an apparent sucde was genune, or f t was murder (n ths case also dsclosng who klled the vctm, how, and why). It uses an archtecture called Moe, desgned to decde, wth the use of adversary search, how and when to gude the user s experence. The nteractve drama s broken down nto abstract peces called user moves, and the system s able to assess any complete sequence of user moves by the evaluaton functon for ts dramatc qualty. A locaton-based pervasve game prototype was developed by [41], to be used durng a car trp, ncludng telephone and walke-talke nteracton. The user plays the role of a detectve (teamed up wth a partner) who tres to uncover an organzed crmnal gang by nvestgatng a seres of crmes that seem to be related. Another prototype system developed to explore locaton-based nteractve stores s Who Klled Hanne Holmgaard? [42], whch s a hstorcal murder mystery set durng World War II. Two partcpants, each playng a dfferent character n the story, should work cooperatvely to unravel the mystery and solve the crme, whle on the move. U-DIRECTOR [43] has a dfferent purpose: to create a drector agent able to orchestrate n real tme the events n a storyworld to mprove the user s experence, copng wth the uncertanty about the user s ntentons and the absence of a 5 Murder, Anyone? (1982). Cncnnat: Vdmax.